feat: package RL CLIs and deployment profile plugins

[wlgys8]

Summary

Closes #131.

• Package training and playback as installed mxtrain / mxplay commands, and export as explicit mxexport onnx / mxexport deploy subcommands. Preserve repository scripts as thin wrappers.
• Ship generic Hydra entry configurations and shared algorithm defaults with motrix_rl; keep concrete task presets in workspace configs/task. Discover launch-directory ./configs by default, honor explicit --config-dir, and prioritize project files over packaged fallbacks.
• Support explicit environment registration modules for train/play without coupling the environment layer to training configuration.
• Discover deployment profile compilers through motrix_deploy.profile_compilers package entry points, keyed by exact environment name. Load only the selected compiler, reject duplicate providers, and remove hardcoded builtin-environment imports from deployment export. Preserve the one-way motrix_rl -> motrix_deploy dependency.
• Migrate bilingual documentation, Docker/demo usage, and tests; remove the unused GPU utility.

Usage

mxtrain task=cartpole/skrl.ppo
mxplay env=cartpole policy=/path/to/policy.pt
mxexport onnx run_dir=/path/to/run output=/tmp/policy.onnx
mxexport deploy run_dir=/path/to/run output=artifacts/policy.deploy
mxexport deploy env=go2-walk-rough

mxexport requires an explicit format. Deployment export selects exactly one of run_dir or env. Installed environment packages declare their profile compilers; external projects do not need to import motrix_envs.

Validation

• RL/provider/workspace suites (not slow and not numba): 504 passed, 2 skipped, 21 deselected.
• Deployment core suite: 255 passed.
• Deployment task + environment-profile suites: 171 passed.
• prek run --all-files and whitespace checks passed.
• Built wheels for motrix_rl, motrix_deploy, and motrix_envs; installed them without dependency replacement into an acceptance environment reusing existing dependencies. Verified command help, packaged configuration loading outside the checkout, and builtin Go2 compiler discovery/build.
• Actual installed mxexport deploy exported a synthetic SKRL-format checkpoint using an external compiler/task package entry point, with an active guard rejecting all motrix_envs imports. Artifact parsing, payload checksums, ONNX checker, and CPU inference passed; four-sample parity maximum absolute error was 1.1920928955078125e-07.
• Actual ONNX export and project-first configuration behavior are covered by integration tests. Independent implementation review found no confirmed actionable defects.

Limitations

• Acceptance reused installed dependencies; it is not a clean dependency-closure installation test.
• Export acceptance used a synthetic checkpoint, not a trained rollout. No GPU training or physical-robot deployment was performed.
• Strict bilingual Sphinx builds could not run because Sphinx is absent; rendered HTML remains unverified.

Base

Rebased onto main after parent PR #130 was squash-merged, so this PR contains only the CLI/configuration/export-plugin migration.

[Copilot]

🔵 Needs a closer look

It spans CLI packaging, a dependency-direction-sensitive entry-point migration, config relocation, and bilingual doc/Docker changes whose packaging and plugin-discovery behavior warrant final human verification.

No open issues 🎉

What changed in this PR

This PR packages the RL control-plane CLIs as installed console scripts (mxtrain, mxplay, mxexport onnx/mxexport deploy) and ships generic Hydra entry configs and shared algo_base defaults inside motrix_rl.configs, keeping concrete task presets in the workspace configs/task. It implements project-first config discovery (launch-directory ./configs, explicit --config-dir, packaged fallback) via a Hydra SearchPathPlugin, decouples environment registration from training config through explicit env_modules, and migrates deployment profile-compiler discovery from an in-process decorator registry to the motrix_deploy.profiles entry-point group keyed by exact environment name. Repository scripts/ become thin wrappers, bilingual docs/Docker/demo usage are updated, and the unused GPU utility is removed.

Changes:

• Add mxtrain/mxplay/mxexport console scripts and package their Hydra configs in motrix_rl.configs, with project-first ./configs discovery and explicit env_modules environment registration.
• Replace the register_profile_compiler decorator registry with motrix_deploy.profiles entry-point discovery (exact-name match, reject duplicates, load only the selected compiler), removing hardcoded motrix_envs imports from export.
• Migrate tests, bilingual documentation, Docker/demo usage, and the DeploymentExportConfig.run → run_dir rename to the new command surface.

File	Description
wiki/​design/​motrix-deploy.md	Documents entry-point profile discovery and the new mxexport deploy entry/dependency direction.
tests/​test_task_configs.py	Composes against packaged motrix_rl.configs with a hydra.searchpath to project configs.
tests/​test_rl_commands.py	New integration tests for installed command composition, project-first config, and help/override behavior.
tests/​test_plugin_export_integration.py	Adds subprocess integration test for mxexport onnx parity outside the workspace.
tests/​test_play_script.py	Imports the packaged motrix_rl.commands.play instead of loading the repo script by path.
tests/​test_doc_generate_video_script.py	Updates expected train command to python -m motrix_rl.commands.train.
tests/​test_deploy_export_cli.py	Uses packaged config module, export_deploy command, and run_dir selector rename.
tests/​test_cli_config.py	Switches train/play/export_onnx composition to the packaged config module with project searchpath.
scripts/​train.py	Reduced to a thin wrapper over the packaged command (truncated in diff).

_{Review effort: Balanced}

[Copilot]

🔵 Needs a closer look

It is a broad cross-package migration altering packaging, entry-point discovery, plugin boundaries, and config schemas across ~90 files, which warrants final human review despite no concrete defects being found.

No open issues 🎉

_{Review effort: Balanced}